AI Found 250,000 New Kinase Substrates From Just 6% Experimental Coverage. The Same Week, It Still Couldn't Reliably Find a Splice Site.

A kinase atlas built from about 20,000 AlphaFold structures turned 6% experimental phosphorylation coverage of the human proteome into roughly 250,000 new, high-confidence kinase-substrate pairs — one of the largest structural leaps AI-for-science has produced this year. Days later, a study of five genomic language models found the same class of network keeps 95 to 100% of its fine-tuned accuracy on promoter detection, a coarse task, but drops to 60 to 88% on splice-site detection, a precise one. A third paper, testing nine general-purpose LLMs on antibody epitopes, names the exact same failure: strong on general biological signal, weak on exact residue-level grounding.

The proteome-scale win: filling in what phosphoproteomics never got to

"An AI-enabled structural atlas decodes kinase specificity across the human proteome" (Vanderwall, Huttlin, Cantley, Gygi and colleagues at Harvard Medical School, Nature Biotechnology, DOI 10.1038/s41587-026-03239-5, published online July 29) starts from a number that undersells the scale of the problem: of the roughly 1.8 million serine, threonine, and tyrosine residues in the human proteome, only about 6% have ever been observed as phosphorylated in an experiment, and only around 5% of those have been mapped to the specific kinase that phosphorylates them. Decades of mass-spec phosphoproteomics has been sampling a small, biased slice of a much larger space.

KinoPlex's approach is to fold in what structure can tell you that a sequence alone can't: it feeds roughly 20,000 AlphaFold-predicted structures into a positive-unlabeled transfer-learning model that asks, for every Ser/Thr/Tyr residue, whether its local 3D environment — solvent exposure, backbone geometry, pocket shape — is compatible with being phosphorylated at all. That step alone flags about 567,000 residues as structurally phospho-competent, already more than five times the experimentally confirmed set. Intersecting those with kinase-specific sequence and structural recognition motifs narrows the list to around 250,000 high-confidence candidates — a specific kinase paired with a specific residue, not just "this site is probably phosphorylated by something."

This is a genuinely different kind of result from a benchmark score. It doesn't claim to have solved kinase specificity; every one of those 250,000 pairs is a candidate for wet-lab or targeted-MS follow-up, not a confirmed fact. What it demonstrates is that structural inference at proteome scale can propose testable biology in a space experiments have barely touched — the kind of leverage a purely sequence-based model doesn't have, because it has no way to reason about whether a residue is physically sitting in a kinase-accessible pocket. Cantley, a senior co-author, has spent decades mapping kinase substrate motifs the slow way, one kinase and one peptide library at a time; the atlas his team built here is explicitly framed as a way to point that kind of follow-up work at the residues most likely to pay off, not as a replacement for it.

The same week: two papers on where the precision breaks down

"Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models" (arXiv 2608.05329, Datta, Shatabda & Rahman, submitted August 5) probes five widely used genomic language models — DNABERT-2, Nucleotide Transformer, HyenaDNA, GENERATOR-v2, and Omni-DNA — by freezing their pretrained weights and testing how much of their fine-tuned performance survives on two different kinds of task. On promoter detection, which asks a model to recognize a broad regulatory region spanning hundreds of base pairs, frozen probes recover 95 to 100% of fine-tuned accuracy: the representation already contains almost everything needed. On splice-site detection, which asks the model to pinpoint the exact base pair where an intron ends and an exon begins, average recovery drops to 60 to 88%. Same models, same pretraining, same freezing procedure — the only thing that changed is whether the task rewards recognizing a region or locating a point.

"EpiBench: Can LLMs Understand Epitopes for Antibody Drug Discovery?" (arXiv 2608.06022, Wang, Wang, Wang, Xu, Du, Hou & Zhang, submitted August 6) runs a version of the same test in a completely different domain. The benchmark spans 1,609 curated samples across five tasks — targetable region discovery, antibody-conditioned epitope identification, epitope binning, functional epitope assessment, and antibody escape assessment — and evaluates nine general-purpose LLMs against it. The paper's own conclusion states the failure mode by name: current LLMs "capture partial epitope-related signals but remain limited in antibody-specific sequence grounding, long-context residue localization, and biologically grounded reasoning." General biological pattern recognition is there. Precisely locating which residues actually form the epitope, on a specific antibody-antigen pair, is not. Two of EpiBench's five tasks make the gap concrete: epitope binning asks whether two antibodies compete for an overlapping surface on the same antigen, and antibody escape assessment asks which mutations would let a virus or tumor antigen slip past a given antibody entirely — both are questions where being approximately right about the neighborhood is functionally equivalent to being wrong, because the therapeutic decision hinges on the exact residue.

Global shape versus exact position

None of these three papers cite each other, and none of the authors set out to make the same point — but line them up and the pattern is hard to miss. KinoPlex succeeds at a task that is fundamentally about compatibility: does this residue's neighborhood look like the kind of place a kinase can reach? That's a coarse-grained, somewhat error-tolerant judgment, and AlphaFold-derived structure gives a model unusually good raw material for making it. Splice-site detection and epitope localization are a different kind of task: they don't ask "is this region compatible," they ask "is it exactly here, and not the base pair or residue next to it." A single wrong position is a wrong answer, not an approximately-right one.

The scale of the two tasks is the tell. A promoter is a few hundred base pairs of regulatory sequence; missing the boundary by ten or twenty base pairs barely changes the call. A splice site is defined by a handful of near-invariant bases at the intron-exon junction — the canonical GT at the start of an intron and AG at its end — and getting that boundary wrong by even one base doesn't produce a slightly-off answer, it produces a different protein or no protein at all. An antibody epitope is the same story at the residue level: two candidates that differ by a single amino acid can bind completely different regions of a target.

That distinction matters because it cuts against a natural but sloppy reading of results like KinoPlex's: if AI can propose 250,000 new kinase-substrate pairs from structure alone, surely the same class of model can pin down a splice site or an epitope residue just as reliably. The three papers this week say no — not because the models are getting worse at biology overall, but because global structural compatibility and exact positional grounding are different problems, solved with different degrees of confidence by the same generation of tools. A 250,000-candidate list is a triage output, not a shortlist ready for a wet lab, and treating it as anything more skips exactly the step these three papers say the field is currently weak at.

What this means for reproducible, local-first science

None of these three papers name or test any commercial research product, MegaBrain BioScience included. What they add up to is a distinction worth having built into how a lab actually works with these outputs: a candidate list and a validated finding are not the same claim, and the gap between them is largest exactly where the task needs a precise position rather than a plausible region. KinoPlex's own funnel makes that gap visible — from 1.8 million residues, to 567,000 structurally plausible, to 250,000 high-confidence pairs, with the wet-lab step still ahead of every one of them.

That's easier to reason about when a researcher can pull a specific candidate out of a list like that, rerun the structural and sequence evidence behind it locally, and see exactly which part of the case is a coarse compatibility call and which part is a precise positional claim — rather than taking a benchmark number or a candidate count at face value. A workbench that runs on your own machine and exports a reproducibility record for every result doesn't change whether a model is good at exact residue grounding. It makes it possible to check, one candidate at a time, whether this one is.